Reproduce “Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tao Zhong, Jonah Buchanan, Christine Allen-Blanchette
Researchers from Princeton University and Lockheed Martin Corporation developed Grasp2Grasp, a vision-based framework that translates dexterous grasp intent between robot hands with different morphologies using Schrödinger Bridges and unpaired grasp datasets. The system achieved a mean grasp success rate of up to 67.45% and a 6D Grasp Wrench Hull IoU of 14.63% when transferring grasps, outperforming existing baselines.
Key insights
- Formulating grasp translation as a Schrödinger Bridge problem enables learning a stochastic process that interpolates between different hand grasp distributions, preserving functional intent.
- Physics-informed ground costs, particularly those derived from grasp wrench space overlap or Jacobian-based manipulability, are critical for achieving stable and functionally aligned grasp transfers without paired data.
- A robust latent space, learned via VAE and incorporating both kinematic and geometric reconstruction, is essential for representing grasp features generalizable across diverse hand morphologies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.